KR100361235B1 - Signal Source Characterization System - Google Patents

Abstract

The present invention starts with linear convolutional mixtures E (t) of primary signals X (t) from sources S1 to Sn, so that the characteristic variable of at least one primary signal X (t) A source characterization system 8 that determines signals F (t) that form estimates. This variable may be an average energy, spectral density, or autocorrelation function. The preprocessing means 20 allows the signals E (t) from the mixtures to be preprocessed before being applied to the signal separating means 10. The measured variable may act to counteract the behavior of the sources.

Application: Control of devices that transmit and / or receive electrical, acoustic or electromagnetic signals

Description

Signal source characterization system

The present invention starts with the first signals E (t) formed by linear convolutive mixing of the primary signals X (t) coming from the respective primary sources, thereby at least one each A source characterization system for supplying at least one second signal for characterizing a primary signal of X (t), wherein the system comprises source separation means.

The invention also relates to the use of such a system for controlling devices for transmitting and / or receiving electrical, acoustic or electromagnetic signals. For example, this may be a car radio or a hands-free telephone.

Background of the Invention

Techniques for separating primary signals from independent sources are known by processing the mixed primary signals to separate and extract each primary signal from these mixed primary signals. This technique applies to primary signals that are only available in the mixed form. This may relate to general linear convolutional blends. These blends can be caused by a variety of causes. These may be caused by the propagation mechanism of the primary signals and / or the overlapping mechanism of signals coming from multiple sources or other causes.

In general, separation techniques are blind. That is, it is assumed that the sources are unknown and independent of the unknown mixed signal. To this end, by detecting multiple samples of these mixtures, one or more original primary signals from these samples can be recovered by a separation algorithm.

Such techniques are known, for example, from the "Blind separation of sources" document published by C. JUTTEN, J. HERAULT in Signal Processing 24 (1991), pages 1-10.

The document describes a source separation means comprising a neural network, which receives the components E (t) of a plurality of signal mixtures at its input and recovers the primary signals separated at its output. This neural network operates cyclically to calculate synaptic coefficients by an adaptive algorithm. In this way, it is possible to process instantaneous linear convolutional mixtures, i.e. at some instant each signal E (t) has a real coefficient which is fixed or slowly changing at this same instant, the primary signals X (t). Is a linear combination of values. The system itself is constantly adapted to changes in these blends. Also, C. JUTTEN and HL NGUYEN THI at Congres Europeen de Mathematiques in Paris, France, July 2-3, 1991, entitled "New algorithms for separation of sources." Known in However, this system has the drawback that it requires substantial computing power for blind separation of all mixed primary sources. This may be a handicap in large devices. It also causes problems with the precision or stability of the sources separated at the output.

Summary of the Invention

It is an object of the present invention to reduce the computational power by taking into account certain characteristics of devices utilizing such source separation.

This object is achieved by a source characterization system comprising a preprocessing means arranged in front of said source separating means, which preprocessing means determines signals E (t) by determining at least one characteristic variable of the first signals E (t). t), which is the third signal I () formed by linear combinations, by fixed or slowly varying coefficients, of characteristic variables of the same nature related to the primary signals X (t). t, p).

By preprocessing the signals E (t) in this way, the source separation means is significantly simplified and processes only linear combinations of characteristic variables of the primary signals, ie instantaneous convolutional linear signal mixtures. In the output, this does not yield separate sources X (t) but rather one or more characteristic variables that characterize these sources. In many uses, in fact, these property variables need only be recognized. If necessary, they can then be processed at the output to obtain separate sources.

This characteristic variable may be an average energy, a spectral density, an autocorrelation function or other variables. This is of great interest, for example, when the output power of a device such as a car radio detects the voice of a hands-free telephone user to enable the processing of the voice information.

Obviously, the same or unequal characteristic variables relating to multiple signals F (t) or multiple primary signals X (t) for one and the same primary signal X (t) The system can also be extended to supply one of a number of signals F (t).

For some characteristic variables, such as the spectral density of the signal, pre-processing can be done prior to the post-processing to convert the calculated characteristic variable into another characteristic variable. According to the invention, the post processing means is arranged at the output of the source separating means. By combining the pre-processing executed by the pre-processing means and the post-processing executed by the post-processing means, it is possible to determine the spectral densities or other characteristic variables for one or more primary signals.

These and other aspects of the invention will be apparent from the embodiments described below.

The present invention may be understood in more detail with reference to the accompanying drawings, which are by no means limiting of the invention.

Detailed description of this embodiment

Different types of signals E (t) are distinguished by the nature of the mixtures from which they originate. Common convolutional linear blends are (1) Generate generic signals E _i (t), where c _ij (t) is the impulse response of the mixed filters, and the symbol ( ^* ) denotes the convolution product.

These general blends include a group of blends propagating with any fixed propagation delay θ and attenuation constant 1 / α _ij . This especially corresponds to the propagation of sound waves outdoors.

The signals in this gen E _i (t) are (2) Form. Where _ij is a constant defining the propagation delay, and the symbol (.) Represents a typical multiplication. These are linear convolutional blends with a fixed non-zero delay.

This family contains mixtures of sub-family in which propagation delay θ _ij is zero. Then, the signals in the genus E _i (t) are (3) Form. Where a _ij is a fixed or slowly changing coefficient (this is different from the coefficient α _ij ). These are linear instantaneous or linear convolutional blends with zero delay.

The above cited documents by C. JUTTEN and J. HERAULT relate to the signals according to equation (3). The invention also relates to two different types of signals.

For example, consider the case where the volume of a car radio installed in a vehicle is controlled. The car radio includes means (not shown) that can be affected by the volume control generated by the car radio in accordance with ambient voice sources. Therefore, if the ambient noise increases (window opening, high speed, driving noise ...), it may be desirable to increase the voice level generated by the car radio. This is not the case when ambient voice sources are formed by the passenger's voice. When passengers are talking, it is a matter of decreasing rather than increasing the volume. This requires identifying passenger voices. 1 shows, for example, passenger voices, various noise sources (engines, bodywork, air circulation through windows, etc.) and primary sources 5 (S1 to Sn) formed by the car radio itself. To identify the voices, transducers Cl-Cn, for example microphones, are arranged inside the passenger compartment. The transducer can pick up the sound emitted by the loudspeaker directly. The microphones detect the first signals E ₁ (t) to E _n (t) from mixtures of primary signals X ₁ (t) to X _n (t) supplied by the sources S1 to Sn. do.

Mixtures formed inside the passenger compartment can be considered as directly related to the propagation of voice signals in the air. In the first approximation, these mixtures can be characterized by the non-zero attenuation coefficient and the non-zero propagation time constant characteristic of each source. The signals E _i (t) detected by the microphone are expressed as follows.

(4)

Where i is the current index of the microphone.

j is the current index of the source.

1 / α _ij , θ _ij are the attenuation coefficients and propagation time constants, ie propagation characteristics from source S _j to microphone E _i , respectively.

According to the invention, the signals E _i (t) are input to a source characterization system 8 comprising a preprocessing means 20 in front of the source separating means 10.

Source separation means capable of separating the instantaneous linear signal mixtures, i.e., mixtures in which all the terms θ _ij are all zero (Equation 3), is much simpler, and thus, non-zero linear convolution according to Equation (2) above. It is easier to realize than the source separation means which can separate the food mixtures. The means for processing linear instantaneous mixing may be the means disclosed in the cited document published by C. JUTTEN and J. HERAULT.

According to the invention, a source separation means 10 behind the preprocessing means 20 and capable of separating linear instantaneous signal mixtures is selected so that a non-zero delay linear formed by the first signals E _i (t) Signal mixtures are transformed to obtain the linear combinations formed by the third signals I _i (t, p). Where t is time and p is an important parameter for the relevant characteristic variable (p is eg frequency). The parameter p is not used if the characteristic variable is the average energy.

The structure of the preprocessing means 20 depends on the characteristic variable of the primary signals X (t) to be identified. This variable is:

The average energy of the primary signal X (t) of one or more sources. In this case, this is the overall characteristic variable determined for a certain duration, ie for successive time intervals,

An autocorrelation function of a plurality of primary signals X (t) or primary signal X (t), determined for a predetermined duration, ie for one or more consecutive time intervals,

May be a spectral density which provides a spectral distribution of the multiple primary signals X (t) or primary signal X (t), determined for a certain duration, ie for one or more consecutive time intervals. .

Those skilled in the art can practice the invention with other characteristic variables without departing from the scope of the invention.

In order to practice the invention, the d.c components of signals E (t) are preferentially removed, for example by low pass filtering. This yields signals E (t) with a mean value of zero, as in the literature of C. JUTTEN and J. HERAULT.

There are certain cases where two sources S1, S2 and two mixed signals E1 (t) and E2 (t) are present, where E1 (t) in either of these is the mixed signal of the sources and the other One is a pure signal (or considered as such) coming directly from one of these sources. In that case, the source separation means 10 can be simplified with an adaptive filtering device that properly weights the signals and then realizes subtraction of the pure signal from the other signals.

This case can be generalized to a pure signal and a signal having a plurality of mixed signals including the pure signal.

Indeed, the pure signal can be measured directly at the output of the autoradio immediately before being transmitted by the loudspeaker.

In this case, for example, the average energy of the primary signal X (t) is considered to be determined. For this purpose, the preprocessing means 20 calculates the average signal energy E _i (t). For this signal E1 (t), this average energy is (5) Is determined for a duration δ starting at the instant t = T ₀ .

In the case of two sources S1, S2 with mixtures according to equation (2), the signal measured by the converter C1 is recorded as follows.

E1 (t) = α ₁₁ .X ₁ (t-θ ₁₁ ) + α ₁₂ X ₂ (t-θ ₁₂ )

The average energy is chosen by selecting δ in such a way that θ ₁₁ , θ _l2 <<δ .. as for uncorrelated sources X ₁ and X ₂ . do.

The delay terms θ ₁₁ , θ ₁₂ disappear and the term ε _E1 is a linear mixture of fixed and slowly varying coefficients and zero delays and the average energies of the sources S1, S2.

Accordingly, the calculation of the average energies of the signals E _i (t) yields signals I (t, p) = ε _Ei (T ₀ , δ) from which the energies of the primary signals can be extracted.

3 shows the element 22 forming part of the pretreatment means 20. This element 22 comprises means 24 for multiplying the signal E _i (t), which is located before the means for storing the multiplication results for a duration δ. At the output, this yields an average energy ε _E1 corresponding to the signal E _i (t).

3 illustrates various digital signal processing. Depending on the analog version the same result can be obtained, for example, by integrating the signal rectified by the capacitances without departing from the scope of the invention.

The preprocessing means 20 comprises a plurality of elements 22 each associated with a signal E _i (t). It is possible to apply time multiplexing. At the output, a number of average energies ε _E1 are obtained, which correspond to each signal E _i (t). All these average energies ε _E1 are input to the source separating means 10 in the manner shown in FIG. 1. They provide output signals F _i (t) which are average energies of each of the primary signals X _i (t) in successive time intervals considered by performing source separation.

Preferably, the source separation means 10 comprises a neural network. This neural network is for example the neural network disclosed in the literature of C. JUTTEN and J HERAULT. According to techniques known to those skilled in the art, the operation of neural networks is

Learning steps to learn to perform the task and

A resolution phase which accordingly uses the learned data to determine the results corresponding to the current values.

These learning techniques are well known and are not described herein. These learning techniques consist in applying samples to the input and modifying the properties of the neural network (basically synaptic coefficients) to generate observed results corresponding to these samples at their output. Using an "unsupervised" learning, ie an algorithm based on these samples, is applied to adapt the synaptic coefficients in such a way that the neural network outputs supply second signals F (t) independent of each other. . Such a learning process is known to those skilled in the art.

In the form of the synaptic coefficients thus determined, the neural network is currently used with new input data. Thus, the results can be determined based on the new input data for the function learned to perform. In order to adapt the system to successive changes of blends to implement the continuous nature of the sources, learning is repeatedly performed to update the synaptic coefficients for the nature of the new blends.

Measured signals E _i (t) corresponding to primary sources X _i (t) with unknown properties are applied to the input, and the means 10 supplies these unknown properties (here, average energies). Supply the measurements of.

The characteristic properties of the primary signals X (t) to be measured are determined at successive time intervals of duration δ. At every hour interval, this

-Single value for average energy,

A series of values obtained by comparing the signal at a moment shifted by a signal and time τ to determine the autocorrelation function,

To determine the spectral densities, it can be related to a number of values determined at different frequencies.

For example, consider the case where the autocorrelation functions Г (τ) are determined for the primary signal X _i (t) with Г _Xi (τ) = ∑X _i (t) .X _i (t + τ). did. To this end, the preprocessing means 20 determines the autocorrelation functions of the signals E _i (t) such that Г _Ei (τ) = ΣE _i (t) .E _i (t + τ). The preprocessing means 20 has preprocessing blocks 25 as shown in FIG. The number of blocks 25 is equal to the number of signals E (t) to be processed. It is also possible to apply to time multiplexing. In order to correlate the signal E (t) with itself, the signal E (t) is applied to the acquiring means 32, which is a continuous sample of the signal E (t) at a predetermined rate (duration D). Take it. This may in particular be a memory having memory elements 32a and 32b in which a series of samples are stored. Shift τ is continuously provided with values D, 2D, 3D, and the like. For example, samples appearing spaced apart by the time D on the outputs 2a, 2b of the memory elements 32a, 32b are applied to the processing means 22 described above, so that the values E (t) .E (t + D). By sliding addressing (arrow 29) of the memory, the value of the autocorrelation function Г _E is determined for the overall relation signal E (t) corresponding to τ = D. The other values of the autocorrelation function are calculated as well as sliding addressing for shifts 2D and 3D, 4D, ... and so on. All these autocorrelation values form an autocorrelation function.

Those skilled in the art will appreciate that the obtaining means 32 may be provided with a structure other than that shown in FIG. 4 without departing from the scope of the present invention. At the output of the element 22, an autocorrelation function Г _E corresponding to the signal E (t) is obtained:

The other signals of FIG. 1 are similarly processed to provide an autocorrelation function Г _E for each signal E (t). In that case, the signals I (t, p) of FIG. 1 are autocorrelation functions Г _E. By applying the same principles as already described for the average energies, the source separating means 10 is trained on the basis of the examples to calculate the autocorrelation functions Г _x corresponding to each of the sources.

This example is limited to only two sources S1 and S2,

Is such that for calculating the auto-correlation function Г _E

E1 (t) = α ₁₁ .X ₁ (t-θ ₁₁ ) + α ₁₂ X ₂ (t-θ ₁₂ )

E2 (t) = α ₂₁ .X ₁ (t-θ ₂₁ ) + α ₂₂ X ₂ (t-θ ₂₂ ) becomes effective.

The source separation means 10 derives the autocorrelation functions Г _x of the source signals such that Г _x (τ) = ΣX (t) .X (t + τ).

When the system 8 determines the autocorrelation functions of the primary signal X (t), it is possible to derive the energies of these signals. To do this, it is sufficient to select the delay tau = 0 in FIG. 4, yielding the same values of Г _x (0) at these outputs.

This mode of operation is faster than the calculation of energies as described above because the source characterization system does the learning in a single time interval rather than in the series of signal intervals. However, this requires more hardware.

In another example, signals I (t, p) may be signals indicating spectral densities. In this case, the preprocessing means 20 is

Calculate the autocorrelation functions and then perform a Fourier transform on each signal E (t) to supply signals I (t, f), where f is the frequency, or

After performing a Fourier transform (e.g. a fast Fourier transform), the modulus square of the Fourier transform determined accordingly is calculated.

All signals I (t, f) derived from signals E (t) and determined for this frequency f are processed in the source separation means 10 as described above. The signals F (t) at the output indicate the spectral densities Dx corresponding to the sources S for the frequency f.

Continuous operation at different frequencies f causes the spectral density function to be obtained for one or more sources by combining the spectrum density values for one source at these different frequencies.

As described above, the output signals of the preprocessing means, which can be average energies, autocorrelation functions, spectral densities, etc., can then be used to influence one or more sources S1-Sn.

Thus, when the average energy of the source S1 is determined in the case of a car radio installed in an automobile, it is possible to control the output amplifier of the car radio according to the measured average energies. Similarly, source detection formed by the speaker's voice may cause the car radio volume to be low or the phone operation to be controlled, for example by sending numbers from the speaker or by enabling the transmission of the voice only when the voice is detected.

If the characteristic variable corresponds to spectral densities, it may affect the equalizer or tone control of the car radio.

The description just above is also valid for other sound reproducing facilities relating to transducers or detectors for signals E (t).

In order to carry out this control, the source characterization system 8 is provided with control means 15 (FIG. 2) for determining the action taken in conjunction with the associated apparatus. Thus, the detection of voice in the passenger space of the vehicle can affect the volume of the car radio. To this end, the control means 15 provide a command 16 that influences the appropriate source.

For other devices, other sources S1-Sn may include signal mixtures caused by radio transmission, such as radio sources. In this case, signals E (t) can be detected, for example, by detectors formed with an antenna. The processing performed by the source characterization system is similar.

In another case, the sources are placed inside the same device. The signals X (t) are delivered to the wires, and the mixtures are formed by cross-talk between these wires. Then, the signals E (t) are not obtained by the converters C1 to Cn.

In order to determine the spectral densities Dx (V), the preprocess 20 described for calculating the autocorrelation functions can be combined with the post process 30 provided at the output of the source separation means 10. 5 relates to this type of processing. The preprocessing means 20 in front of the source separating means 10 determines the autocorrelation functions Г _Xi (t) and the post processing means 30 is arranged at the output applying the Fourier transform.

The present invention has been described for a linear convolutional signal comprising gain terms and pure delay terms defined in equation (2).

The invention can also be applied in general to convolutional mixtures, ie mixtures defined by formula (1).

In order to reduce the general case to the specific case of equation (2) described above, this process is applied to signals E (t) in which the frequency band is preferentially limited. To this end, the preprocessing means 20 comprises identification filters 40 (FIG. 6) at the input, which limit each signal Ei (t) to a smaller frequency band. The processing remains as described above but the results obtained thus relate not only to the primary signals X (t) but also to the filtered version of their mixtures. Nevertheless, these results are adapted to the source properties even in the case of any convolutional blends. For example, when average energies or spectral densities are calculated, signals relating to successive frequency bands can be combined at the output to obtain the above characteristic variables for the entire frequency spectrum.

1 is a block diagram of a source characterization system including pre-processing means.

2 shows a source characterization system comprising control means which act in opposition to primary signal sources.

3 is an example of a diagram of a portion of said preprocessing means for determining an average energy of a signal.

4 is an example of a diagram of a portion of said preprocessing means for determining autocorrelation functions.

5 is a block diagram of a source characterization system comprising pre-processing means and post-processing means.

FIG. 6 is a view similar to that shown in FIG. 3 with filter means for processing common convolutive mixtures.

Explanation of symbols on the main parts of the drawings

5: Primary Source 8: Source Characterization System

10 source separation means 20 pretreatment means

Claims (7)

Starting from the first signals E (t) formed by linear convolutional mixing of the primary signals X (t) of the respective primary sources 5, at least one respective primary signal X (t) In a source characterization system (8) comprising at least one second signal (F) for characterizing and comprising a source separation means (10), A preprocessing means 20 arranged in front of said source separating means 10, said preprocessing means preprocessing said signals E (t) by determining at least one characteristic variable of said signals E (t). Wherein said variable is the third signals I (t, p) formed by linear combinations, by fixed or slowly varying coefficients of characteristic variables of the same nature related to said primary signals X (t). Source characterization system characterized in that it is supplied in the form of. The method of claim 1, Wherein said characteristic variable is an average energy, spectral density and autocorrelation function. The method according to claim 1 or 2, A post processing means 30 disposed at the output of the source separating means 10 for post processing the fourth signals G (t) supplied by the source separating means 10, the fixed Said linear combinations with coefficients are determined by combining said post-process and said pre-process. The method of claim 3, wherein And said post-processing means supplying estimates of primary signals X (t). The method according to claim 1 or 2, Control means 15 for receiving at the input second signals F, based on which said control means supplies at least one command 16 which affects at least one source. Source characterization system. A system applying the source characterization system of claim 1 for controlling devices transmitting and / or receiving electrical, acoustic or electromagnetic signals. The method of claim 6, A system employing a source characterization system for controlling the sound generated by a car radio or for controlling hands-free telephone operation.